llm-wiki
This isn't a tool you install — it's a 2-page gist from Andrej Karpathy (April 2026) describing a pattern: instead of RAG-ing over messy documents, have the LLM build and maintain a structured wiki of interlinked Markdown pages as it works. Every answer is a lookup or an edit. Every edit is citable. The pattern is the thing most new memory tools in 2026 are implementing.
- Anyone orienting themselves to where AI memory is going in 2026
- Builders picking an implementation — you should read this first
- CEOs who want the conceptual vocabulary to reason about memory tooling
What you'll do
llm-wiki is a pattern, not software. Read the gist (it's about 2 pages), internalize the mental model, then pick a concrete implementation below — gbrain, claude-memory-compiler, or obsidian-wiki are the current serious options.
Before you start
- ~5 minutes of reading time
Step-by-step install
- 011. Read the gist
Open the original gist at gist.github.com/karpathy. It's short. The core claim: instead of RAG over raw documents, have the LLM compile a structured wiki of interlinked Markdown pages about everything it knows. Ask questions → it reads pages. Learn something new → it edits pages. Each page cross-references others.
- 022. Understand why it matters
RAG gets better at retrieval over time. But the underlying docs stay messy. The wiki pattern instead gets better at the *substrate* — the knowledge base itself becomes more organized and more navigable. Your agent compiles institutional knowledge the way a good Wikipedia editor would.
- 033. Pick an implementation
The serious implementations as of April 2026: gbrain (Markdown + pgvector, Garry Tan's), claude-memory-compiler (Claude Code hooks, Cole Medin's), obsidian-wiki (Markdown framework, Ar9av's). Click any tool card in the registry for its install guide.
Your first 10 minutes
- 01Skim the gist once more with your own org in mind. What would the first 5 wiki pages be?
- 02Pick the implementation closest to your tech fluency — obsidian-wiki if you already use Obsidian, claude-memory-compiler if you live in Claude Code, gbrain if you want the most production-proven version.
- 03Install your chosen implementation (each has its own guide).
- 04Seed the wiki with 5 pages about your company: mission, customers, team, priorities, recent decisions.
- 05Layer Cognition CLO on top to track which concepts your team is retaining.
Troubleshooting
I read the gist and I'm overwhelmed.
That's normal. Start with obsidian-wiki or obsidian-mcp-tools for the lowest-friction way to see the pattern in action. The gist clicks after you've touched a concrete implementation.
llm-wiki holds the knowledge. Cognition CLO models retention per employee per concept using a Weibull forgetting curve — so you see decay before it becomes a missed SOP or a failed audit.